2025 Proffered Presentations
S238: PREDICTING PITUITARY ADENOMA CONSISTENCY USING RADIOMIC DATA MINING AND MACHINE LEARNING ON T2-WEIGHTED MRI: A MULTICENTER RETROSPECTIVE STUDY
Edoardo Agosti1; Lorenzo Ugga2; Marcello Mangili1; Renato Cuocolo2; Vittorio Rampinelli1; Pierlorenzo Veiceschi3; Marina Cappelletti4; Pier Paolo Panciani1; Luigi Maria Cavallo2; Davide Locatelli5; Alessandro Fiorindi1; Marco Maria Fontanella1; 1University of Brescia; 2Federico II University; 3Ospedale Civico of Palermo; 4Ospedale Ca' Forcello of Treviso; 5Insubria University
Introduction: The ease of removing a pituitary adenoma (PA) during surgery, particularly with a transsphenoidal approach, can be affected by the tumor's consistency. Standard qualitative magnetic resonance imaging (MRI) does not allow for the assessment of this consistency. However, radiomic texture analysis may provide quantitative tissue characteristics. This study aimed to assess the accuracy of using texture analysis along with machine learning techniques for the preoperative evaluation of PA consistency in patients scheduled for endonasal endoscopic surgery.
Methods: Data from patients with PA who underwent transsphenoidal surgery at the centers in Brescia, Naples, Varese, and Treviso were retrospectively reviewed. Pituitary adenomas (PAs) were classified based on intraoperative findings as soft, fibrous, or mixed (if they exhibited characteristics of both). Specifically, PAs that could be easily removed using standard maneuvers like curettage and suction were defined as soft. Those that were more resistant, difficult to remove, and required more complex procedures such as extracapsular dissection have been classified as fibrous. All patients underwent MRI exams either on a 1.5 or 3 T MR scanner. The imaging protocol always included a coronal T2-weighted Turbo Spin Echo sequence. Manual segmentation was performed with ITK-Snap software on T2-weighted sequence (Figure 1). After manual segmentation, radiomic texture features were extracted from the original and filtered MRI images. Stability analysis of features and multi-stage feature selection were performed. After oversampling to balance the classes, 80% of the data was used for hyperparameter optimization through 5-fold stratified cross-validation, while a 20% hold-out set was used for final testing, employing an Extra Trees ensemble meta-algorithm. The reference standard was based on surgical outcomes.
Results: A total of 400 patients were collected, of with 210 had a soft PA, 152 a mixed PA, and 38 a fibrous PA. A total of 1003 texture features were extracted, of which 678 were stable. After removing low-variance parameters (n=5) and highly intercorrelated features (n=591), recursive feature elimination identified a subset of 13 features. After hyperparameter tuning, the Extra Trees classifier achieved an accuracy of 94%, sensitivity of 100%, and specificity of 88%. The area under the receiver operating characteristic curve and precision-recall curves was 0.99.
Conclusion: The machine learning model, trained on radiomic features from T2-weighted MRI, can efficiently aid in distinguishing between soft and fibrous PAs preoperatively. Consequently, with further development and validation on larger datasets, this tool could be highly beneficial for pre-surgical planning in these patients.
